├── .github └── workflows │ ├── cla.yml │ ├── publish.yml │ └── tests.yml ├── .gitignore ├── CLA.md ├── LICENSE ├── README.md ├── benchmarks ├── benchmark.py └── scoring.py ├── extract.py ├── poetry.lock ├── pyproject.toml ├── run_table_app.py ├── scripts └── verify_benchmark_scores.py ├── static └── images │ └── table_example.png ├── table_app.py └── tabled ├── assignment.py ├── extract.py ├── fileinput.py ├── formats ├── __init__.py ├── common.py ├── csv.py ├── html.py └── markdown.py ├── heuristics ├── __init__.py └── cells.py ├── inference ├── detection.py ├── models.py └── recognition.py ├── schema.py └── settings.py /.github/workflows/cla.yml: -------------------------------------------------------------------------------- 1 | name: "Tabled CLA Assistant" 2 | on: 3 | issue_comment: 4 | types: [created] 5 | pull_request_target: 6 | types: [opened,closed,synchronize] 7 | 8 | # explicitly configure permissions, in case your GITHUB_TOKEN workflow permissions are set to read-only in repository settings 9 | permissions: 10 | actions: write 11 | contents: write 12 | pull-requests: write 13 | statuses: write 14 | 15 | jobs: 16 | CLAAssistant: 17 | runs-on: ubuntu-latest 18 | steps: 19 | - name: "Tabled CLA Assistant" 20 | if: (github.event.comment.body == 'recheck' || github.event.comment.body == 'I have read the CLA Document and I hereby sign the CLA') || github.event_name == 'pull_request_target' 21 | uses: contributor-assistant/github-action@v2.3.0 22 | env: 23 | GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} 24 | # the below token should have repo scope and must be manually added by you in the repository's secret 25 | # This token is required only if you have configured to store the signatures in a remote repository/organization 26 | PERSONAL_ACCESS_TOKEN: ${{ secrets.PERSONAL_ACCESS_TOKEN }} 27 | with: 28 | path-to-signatures: 'signatures/version1/cla.json' 29 | path-to-document: 'https://github.com/VikParuchuri/tabled/blob/master/CLA.md' 30 | # branch should not be protected 31 | branch: 'master' 32 | allowlist: VikParuchuri -------------------------------------------------------------------------------- /.github/workflows/publish.yml: -------------------------------------------------------------------------------- 1 | name: Python package 2 | on: 3 | push: 4 | tags: 5 | - "v*.*.*" 6 | jobs: 7 | build: 8 | runs-on: ubuntu-latest 9 | steps: 10 | - uses: actions/checkout@v3 11 | - name: Set up Python 3.11 12 | uses: actions/setup-python@v4 13 | with: 14 | python-version: 3.11 15 | - name: Install python dependencies 16 | run: | 17 | pip install poetry 18 | poetry install 19 | - name: Build package 20 | run: | 21 | poetry build 22 | - name: Publish package 23 | env: 24 | PYPI_TOKEN: ${{ secrets.PYPI_TOKEN }} 25 | run: | 26 | poetry config pypi-token.pypi "$PYPI_TOKEN" 27 | poetry publish 28 | -------------------------------------------------------------------------------- /.github/workflows/tests.yml: -------------------------------------------------------------------------------- 1 | name: Integration test 2 | 3 | on: [push] 4 | 5 | env: 6 | TORCH_DEVICE: "cpu" 7 | 8 | jobs: 9 | build: 10 | runs-on: ubuntu-latest 11 | steps: 12 | - uses: actions/checkout@v3 13 | - name: Set up Python 3.11 14 | uses: actions/setup-python@v4 15 | with: 16 | python-version: 3.11 17 | - name: Install apt dependencies 18 | run: | 19 | sudo apt-get update 20 | - name: Install python dependencies 21 | run: | 22 | pip install poetry 23 | poetry install 24 | poetry run pip uninstall torch -y 25 | poetry run pip install torch --index-url https://download.pytorch.org/whl/cpu 26 | - name: Run benchmark test 27 | run: | 28 | poetry run python benchmarks/benchmark.py --max 5 temp.json 29 | poetry run python scripts/verify_benchmark_scores.py temp.json -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | private.py 2 | .DS_Store 3 | local.env 4 | notebooks 5 | results 6 | 7 | # Byte-compiled / optimized / DLL files 8 | __pycache__/ 9 | *.py[cod] 10 | *$py.class 11 | 12 | # C extensions 13 | *.so 14 | 15 | # Distribution / packaging 16 | .Python 17 | build/ 18 | develop-eggs/ 19 | dist/ 20 | downloads/ 21 | eggs/ 22 | .eggs/ 23 | lib/ 24 | lib64/ 25 | parts/ 26 | sdist/ 27 | var/ 28 | wheels/ 29 | share/python-wheels/ 30 | *.egg-info/ 31 | .installed.cfg 32 | *.egg 33 | MANIFEST 34 | 35 | # PyInstaller 36 | # Usually these files are written by a python script from a template 37 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 38 | *.manifest 39 | *.spec 40 | 41 | # Installer logs 42 | pip-log.txt 43 | pip-delete-this-directory.txt 44 | 45 | # Unit test / coverage reports 46 | htmlcov/ 47 | .tox/ 48 | .nox/ 49 | .coverage 50 | .coverage.* 51 | .cache 52 | nosetests.xml 53 | coverage.xml 54 | *.cover 55 | *.py,cover 56 | .hypothesis/ 57 | .pytest_cache/ 58 | cover/ 59 | 60 | # Translations 61 | *.mo 62 | *.pot 63 | 64 | # Django stuff: 65 | *.log 66 | local_settings.py 67 | db.sqlite3 68 | db.sqlite3-journal 69 | 70 | # Flask stuff: 71 | instance/ 72 | .webassets-cache 73 | 74 | # Scrapy stuff: 75 | .scrapy 76 | 77 | # Sphinx documentation 78 | docs/_build/ 79 | 80 | # PyBuilder 81 | .pybuilder/ 82 | target/ 83 | 84 | # Jupyter Notebook 85 | .ipynb_checkpoints 86 | 87 | # IPython 88 | profile_default/ 89 | ipython_config.py 90 | 91 | # pyenv 92 | # For a library or package, you might want to ignore these files since the code is 93 | # intended to run in multiple environments; otherwise, check them in: 94 | # .python-version 95 | 96 | # pipenv 97 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 98 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 99 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 100 | # install all needed dependencies. 101 | #Pipfile.lock 102 | 103 | # poetry 104 | # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. 105 | # This is especially recommended for binary packages to ensure reproducibility, and is more 106 | # commonly ignored for libraries. 107 | # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control 108 | #poetry.lock 109 | 110 | # pdm 111 | # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. 112 | #pdm.lock 113 | # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it 114 | # in version control. 115 | # https://pdm.fming.dev/#use-with-ide 116 | .pdm.toml 117 | 118 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm 119 | __pypackages__/ 120 | 121 | # Celery stuff 122 | celerybeat-schedule 123 | celerybeat.pid 124 | 125 | # SageMath parsed files 126 | *.sage.py 127 | 128 | # Environments 129 | .env 130 | .venv 131 | env/ 132 | venv/ 133 | ENV/ 134 | env.bak/ 135 | venv.bak/ 136 | 137 | # Spyder project settings 138 | .spyderproject 139 | .spyproject 140 | 141 | # Rope project settings 142 | .ropeproject 143 | 144 | # mkdocs documentation 145 | /site 146 | 147 | # mypy 148 | .mypy_cache/ 149 | .dmypy.json 150 | dmypy.json 151 | 152 | # Pyre type checker 153 | .pyre/ 154 | 155 | # pytype static type analyzer 156 | .pytype/ 157 | 158 | # Cython debug symbols 159 | cython_debug/ 160 | 161 | # PyCharm 162 | # JetBrains specific template is maintained in a separate JetBrains.gitignore that can 163 | # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore 164 | # and can be added to the global gitignore or merged into this file. For a more nuclear 165 | # option (not recommended) you can uncomment the following to ignore the entire idea folder. 166 | .idea/ -------------------------------------------------------------------------------- /CLA.md: -------------------------------------------------------------------------------- 1 | Tabled Contributor Agreement 2 | 3 | This Tabled Contributor Agreement ("TCA") applies to any contribution that you make to any product or project managed by us (the "project"), and sets out the intellectual property rights you grant to us in the contributed materials. The term "us" shall mean Endless Labs, Inc. The term "you" shall mean the person or entity identified below. 4 | 5 | If you agree to be bound by these terms, sign by writing "I have read the CLA document and I hereby sign the CLA" in response to the CLA bot Github comment. Read this agreement carefully before signing. These terms and conditions constitute a binding legal agreement. 6 | 7 | 1. The term 'contribution' or 'contributed materials' means any source code, object code, patch, tool, sample, graphic, specification, manual, documentation, or any other material posted or submitted by you to the project. 8 | 2. With respect to any worldwide copyrights, or copyright applications and registrations, in your contribution: 9 | - you hereby assign to us joint ownership, and to the extent that such assignment is or becomes invalid, ineffective or unenforceable, you hereby grant to us a perpetual, irrevocable, non-exclusive, worldwide, no-charge, royalty free, unrestricted license to exercise all rights under those copyrights. This includes, at our option, the right to sublicense these same rights to third parties through multiple levels of sublicensees or other licensing arrangements, including dual-license structures for commercial customers; 10 | - you agree that each of us can do all things in relation to your contribution as if each of us were the sole owners, and if one of us makes a derivative work of your contribution, the one who makes the derivative work (or has it made will be the sole owner of that derivative work; 11 | - you agree that you will not assert any moral rights in your contribution against us, our licensees or transferees; 12 | - you agree that we may register a copyright in your contribution and exercise all ownership rights associated with it; and 13 | - you agree that neither of us has any duty to consult with, obtain the consent of, pay or render an accounting to the other for any use or distribution of vour contribution. 14 | 3. With respect to any patents you own, or that you can license without payment to any third party, you hereby grant to us a perpetual, irrevocable, non-exclusive, worldwide, no-charge, royalty-free license to: 15 | - make, have made, use, sell, offer to sell, import, and otherwise transfer your contribution in whole or in part, alone or in combination with or included in any product, work or materials arising out of the project to which your contribution was submitted, and 16 | - at our option, to sublicense these same rights to third parties through multiple levels of sublicensees or other licensing arrangements. 17 | If you or your affiliates institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the contribution or any project it was submitted to constitutes direct or contributory patent infringement, then any patent licenses granted to you under this agreement for that contribution shall terminate as of the date such litigation is filed. 18 | 4. Except as set out above, you keep all right, title, and interest in your contribution. The rights that you grant to us under these terms are effective on the date you first submitted a contribution to us, even if your submission took place before the date you sign these terms. Any contribution we make available under any license will also be made available under a suitable FSF (Free Software Foundation) or OSI (Open Source Initiative) approved license. 19 | 5. You covenant, represent, warrant and agree that: 20 | - each contribution that you submit is and shall be an original work of authorship and you can legally grant the rights set out in this SCA; 21 | - to the best of your knowledge, each contribution will not violate any third party's copyrights, trademarks, patents, or other intellectual property rights; and 22 | - each contribution shall be in compliance with U.S. export control laws and other applicable export and import laws. 23 | You agree to notify us if you become aware of any circumstance which would make any of the foregoing representations inaccurate in any respect. Endless Labs, Inc. may publicly disclose your participation in the project, including the fact that you have signed the TCA. 24 | 6. This TCA is governed by the laws of the State of California and applicable U.S. Federal law. Any choice of law rules will not apply. -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | Tabled table extractor 635 | Copyright (C) 2024 Vikas Paruchuri 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Tabled Copyright (C) 2024 Vikas Paruchuri 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | > [!IMPORTANT] 2 | > Tabled is now deprecated. The functionality here has been migrated to [marker](https://github.com/vikParuchuri/marker). 3 | > To extract tables from a PDF with marker, you can run `python convert_single.py FILENAME --converter_cls marker.converters.table.TableConverter` 4 | > Read more, and see other CLI options [here](https://github.com/vikParuchuri/marker?tab=readme-ov-file#extract-tables). 5 | 6 | # Tabled 7 | 8 | Tabled is a small library for detecting and extracting tables. It uses [surya](https://www.github.com/VikParuchuri/surya) to find all the tables in a PDF, identifies the rows/columns, and formats cells into markdown, csv, or html. 9 | 10 | ## Example 11 | 12 | ![Table image 0](static/images/table_example.png) 13 | 14 | 15 | | Characteristic | | | Population | | | | Change from 2016 to 2060 | | 16 | |--------------------|-------|-------|--------------|-------|-------|-------|------------------------------|---------| 17 | | | 2016 | 2020 | 2030 | 2040 | 2050 | 2060 | Number | Percent | 18 | | Total population | 323.1 | 332.6 | 355.1 | 373.5 | 388.9 | 404.5 | 81.4 | 25.2 | 19 | | Under 18 years | 73.6 | 74.0 | 75.7 | 77.1 | 78.2 | 80.1 | 6.5 | 8.8 | 20 | | 18 to 44 years | 116.0 | 119.2 | 125.0 | 126.4 | 129.6 | 132.7 | 16.7 | 14.4 | 21 | | 45 to 64 years | 84.3 | 83.4 | 81.3 | 89.1 | 95.4 | 97.0 | 12.7 | 15.1 | 22 | | 65 years and over | 49.2 | 56.1 | 73.1 | 80.8 | 85.7 | 94.7 | 45.4 | 92.3 | 23 | | 85 years and over | 6.4 | 6.7 | 9.1 | 14.4 | 18.6 | 19.0 | 12.6 | 198.1 | 24 | | 100 years and over | 0.1 | 0.1 | 0.1 | 0.2 | 0.4 | 0.6 | 0.5 | 618.3 | 25 | 26 | 27 | ## Community 28 | 29 | [Discord](https://discord.gg//KuZwXNGnfH) is where we discuss future development. 30 | 31 | # Hosted API 32 | 33 | There is a hosted API for tabled available [here](https://www.datalab.to/): 34 | 35 | - Works with PDF, images, word docs, and powerpoints 36 | - Consistent speed, with no latency spikes 37 | - High reliability and uptime 38 | 39 | # Commercial usage 40 | 41 | I want tabled to be as widely accessible as possible, while still funding my development/training costs. Research and personal usage is always okay, but there are some restrictions on commercial usage. 42 | 43 | The weights for the models are licensed `cc-by-nc-sa-4.0`, but I will waive that for any organization under $5M USD in gross revenue in the most recent 12-month period AND under $5M in lifetime VC/angel funding raised. You also must not be competitive with the [Datalab API](https://www.datalab.to/). If you want to remove the GPL license requirements (dual-license) and/or use the weights commercially over the revenue limit, check out the options [here](https://www.datalab.to). 44 | 45 | # Installation 46 | 47 | You'll need python 3.10+ and PyTorch. You may need to install the CPU version of torch first if you're not using a Mac or a GPU machine. See [here](https://pytorch.org/get-started/locally/) for more details. 48 | 49 | Install with: 50 | 51 | ```shell 52 | pip install tabled-pdf 53 | ``` 54 | 55 | Post-install: 56 | 57 | - Inspect the settings in `tabled/settings.py`. You can override any settings with environment variables. 58 | - Your torch device will be automatically detected, but you can override this. For example, `TORCH_DEVICE=cuda`. 59 | - Model weights will automatically download the first time you run tabled. 60 | 61 | # Usage 62 | 63 | ```shell 64 | tabled DATA_PATH 65 | ``` 66 | 67 | - `DATA_PATH` can be an image, pdf, or folder of images/pdfs 68 | - `--format` specifies output format for each table (`markdown`, `html`, or `csv`) 69 | - `--save_json` saves additional row and column information in a json file 70 | - `--save_debug_images` saves images showing the detected rows and columns 71 | - `--skip_detection` means that the images you pass in are all cropped tables and don't need any table detection. 72 | - `--detect_cell_boxes` by default, tabled will attempt to pull cell information out of the pdf. If you instead want cells to be detected by a detection model, specify this (usually you only need this with pdfs that have bad embedded text). 73 | - `--save_images` specifies that images of detected rows/columns and cells should be saved. 74 | 75 | After running the script, the output directory will contain folders with the same basenames as the input filenames. Inside those folders will be the markdown files for each table in the source documents. There will also optionally be images of the tables. 76 | 77 | There will also be a `results.json` file in the root of the output directory. The file will contain a json dictionary where the keys are the input filenames without extensions. Each value will be a list of dictionaries, one per table in the document. Each table dictionary contains: 78 | 79 | - `cells` - the detected text and bounding boxes for each table cell. 80 | - `bbox` - bbox of the cell within the table bbox 81 | - `text` - the text of the cell 82 | - `row_ids` - ids of rows the cell belongs to 83 | - `col_ids` - ids of columns the cell belongs to 84 | - `order` - order of this cell within its assigned row/column cell. (sort by row, then column, then order) 85 | - `rows` - bboxes of the detected rows 86 | - `bbox` - bbox of the row in (x1, x2, y1, y2) format 87 | - `row_id` - unique id of the row 88 | - `cols` - bboxes of detected columns 89 | - `bbox` - bbox of the column in (x1, x2, y1, y2) format 90 | - `col_id` - unique id of the column 91 | - `image_bbox` - the bbox for the image in (x1, y1, x2, y2) format. (x1, y1) is the top left corner, and (x2, y2) is the bottom right corner. The table bbox is relative to this. 92 | - `bbox` - the bounding box of the table within the image bbox. 93 | - `pnum` - page number within the document 94 | - `tnum` - table index on the page 95 | 96 | ## Interactive App 97 | 98 | I've included a streamlit app that lets you interactively try tabled on images or PDF files. Run it with: 99 | 100 | ```shell 101 | pip install streamlit 102 | tabled_gui 103 | ``` 104 | 105 | ## From python 106 | 107 | ```python 108 | from tabled.extract import extract_tables 109 | from tabled.fileinput import load_pdfs_images 110 | from tabled.inference.models import load_detection_models, load_recognition_models, load_layout_models 111 | 112 | det_models, rec_models, layout_models = load_detection_models(), load_recognition_models(), load_layout_models() 113 | images, highres_images, names, text_lines = load_pdfs_images(IN_PATH) 114 | 115 | page_results = extract_tables(images, highres_images, text_lines, det_models, layout_models, rec_models) 116 | ``` 117 | 118 | # Benchmarks 119 | 120 | | Avg score | Time per table | Total tables | 121 | |-------------|------------------|----------------| 122 | | 0.847 | 0.029 | 688 | 123 | 124 | ## Quality 125 | 126 | Getting good ground truth data for tables is hard, since you're either constrained to simple layouts that can be heuristically parsed and rendered, or you need to use LLMs, which make mistakes. I chose to use GPT-4 table predictions as a pseudo-ground-truth. 127 | 128 | Tabled gets a `.847` alignment score when compared to GPT-4, which indicates alignment between the text in table rows/cells. Some of the misalignments are due to GPT-4 mistakes, or small inconsistencies in what GPT-4 considered the borders of the table. In general, extraction quality is quite high. 129 | 130 | ## Performance 131 | 132 | Running on an A10G with 10GB of VRAM usage and batch size `64`, tabled takes `.029` seconds per table. 133 | 134 | ## Running the benchmark 135 | 136 | Run the benchmark with: 137 | 138 | ```shell 139 | python benchmarks/benchmark.py out.json 140 | ``` 141 | 142 | # Acknowledgements 143 | 144 | - Thank you to [Peter Jansen](https://cognitiveai.org/) for the benchmarking dataset, and for discussion about table parsing. 145 | - Huggingface for inference code and model hosting 146 | - PyTorch for training/inference -------------------------------------------------------------------------------- /benchmarks/benchmark.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import json 3 | import time 4 | 5 | import click 6 | import datasets 7 | from surya.input.pdflines import get_table_blocks 8 | from tabulate import tabulate 9 | from tqdm import tqdm 10 | from scoring import score_table 11 | from tabled.assignment import assign_rows_columns 12 | 13 | from tabled.formats import formatter 14 | from tabled.inference.models import load_recognition_models 15 | from tabled.inference.recognition import recognize_tables 16 | 17 | 18 | @click.command() 19 | @click.argument("out_file", type=str) 20 | @click.option("--dataset", type=str, default="vikp/table_bench2", help="Dataset to use") 21 | @click.option("--max", type=int, default=None, help="Max number of tables to process") 22 | def main(out_file, dataset, max): 23 | ds = datasets.load_dataset(dataset, split="train") 24 | 25 | rec_models = load_recognition_models() 26 | 27 | results = [] 28 | table_imgs = [] 29 | table_blocks = [] 30 | image_sizes = [] 31 | iterations = len(ds) 32 | if max is not None: 33 | iterations = min(max, len(ds)) 34 | for i in range(iterations): 35 | row = ds[i] 36 | line_data = json.loads(row["text_lines"]) 37 | table_bbox = row["table_bbox"] 38 | image_size = row["page_size"] 39 | table_img = row["table_image"] 40 | 41 | table_block = get_table_blocks([table_bbox], line_data, image_size)[0] 42 | table_imgs.append(table_img) 43 | table_blocks.append(table_block) 44 | image_sizes.append(image_size) 45 | 46 | start = time.time() 47 | table_rec = recognize_tables(table_imgs, table_blocks, [False] * len(table_imgs), rec_models) 48 | total_time = time.time() - start 49 | cells = [assign_rows_columns(tr, im_size) for tr, im_size in zip(table_rec, image_sizes)] 50 | 51 | for i in range(iterations): 52 | row = ds[i] 53 | table_cells = cells[i] 54 | table_bbox = row["table_bbox"] 55 | gpt4_table = json.loads(row["gpt_4_table"])["markdown_table"] 56 | 57 | table_markdown, _ = formatter("markdown", table_cells) 58 | 59 | results.append({ 60 | "score": score_table(table_markdown, gpt4_table), 61 | "arxiv_id": row["arxiv_id"], 62 | "page_idx": row["page_idx"], 63 | "marker_table": table_markdown, 64 | "gpt4_table": gpt4_table, 65 | "table_bbox": table_bbox 66 | }) 67 | 68 | avg_score = sum([r["score"] for r in results]) / len(results) 69 | headers = ["Avg score", "Time per table", "Total tables"] 70 | data = [f"{avg_score:.3f}", f"{total_time / len(ds):.3f}", len(ds)] 71 | 72 | table = tabulate([data], headers=headers, tablefmt="github") 73 | print(table) 74 | print("Avg score computed by aligning table cell text with GPT-4 table cell text.") 75 | 76 | with open(out_file, "w+") as f: 77 | json.dump(results, f, indent=2) 78 | 79 | 80 | if __name__ == "__main__": 81 | main() -------------------------------------------------------------------------------- /benchmarks/scoring.py: -------------------------------------------------------------------------------- 1 | from rapidfuzz import fuzz 2 | import re 3 | 4 | 5 | def split_to_cells(table): 6 | table = table.strip() 7 | table = re.sub(r" {2,}", "", table) 8 | table_rows = table.split("\n") 9 | table_rows = [t for t in table_rows if t.strip()] 10 | table_cells = [[c.strip() for c in r.split("|")] for r in table_rows] 11 | return table_cells 12 | 13 | 14 | def align_rows(hypothesis, ref_row): 15 | best_alignment = [] 16 | best_alignment_score = 0 17 | for j in range(0, len(hypothesis)): 18 | alignments = [] 19 | for i in range(len(ref_row)): 20 | if i >= len(hypothesis[j]): 21 | alignments.append(0) 22 | continue 23 | alignment = fuzz.ratio(hypothesis[j][i], ref_row[i], score_cutoff=30) / 100 24 | alignments.append(alignment) 25 | if len(alignments) == 0: 26 | continue 27 | alignment_score = sum(alignments) / len(alignments) 28 | if alignment_score >= best_alignment_score: 29 | best_alignment = alignments 30 | best_alignment_score = alignment_score 31 | return best_alignment 32 | 33 | 34 | def score_table(hypothesis, reference): 35 | hypothesis = split_to_cells(hypothesis) 36 | reference = split_to_cells(reference) 37 | 38 | alignments = [] 39 | for i in range(0, len(reference)): 40 | alignments.extend(align_rows(hypothesis, reference[i])) 41 | return sum(alignments) / max(len(alignments), 1) -------------------------------------------------------------------------------- /extract.py: -------------------------------------------------------------------------------- 1 | import json 2 | from collections import defaultdict 3 | 4 | import copy 5 | import os 6 | 7 | import click 8 | from surya.postprocessing.heatmap import draw_bboxes_on_image 9 | 10 | from tabled.extract import extract_tables 11 | from tabled.formats import formatter 12 | from tabled.fileinput import load_pdfs_images 13 | from tabled.inference.models import load_detection_models, load_recognition_models, load_layout_models 14 | 15 | 16 | @click.command(help="Extract tables from PDFs") 17 | @click.argument("in_path", type=click.Path(exists=True)) 18 | @click.argument("out_folder", type=click.Path()) 19 | @click.option("--save_json", is_flag=True, help="Save row/column/cell information in json format") 20 | @click.option("--save_debug_images", is_flag=True, help="Save images for debugging") 21 | @click.option("--skip_detection", is_flag=True, help="Skip table detection") 22 | @click.option("--detect_cell_boxes", is_flag=True, help="Detect table cell boxes vs extract from PDF. Will also run OCR.") 23 | @click.option("--format", type=click.Choice(["markdown", "csv", "html"]), default="markdown") 24 | def main(in_path, out_folder, save_json, save_debug_images, skip_detection, detect_cell_boxes, format): 25 | os.makedirs(out_folder, exist_ok=True) 26 | images, highres_images, names, text_lines = load_pdfs_images(in_path) 27 | pnums = [] 28 | prev_name = None 29 | for i, name in enumerate(names): 30 | if prev_name is None or prev_name != name: 31 | pnums.append(0) 32 | else: 33 | pnums.append(pnums[-1] + 1) 34 | 35 | prev_name = name 36 | 37 | det_models = load_detection_models() 38 | rec_models = load_recognition_models() 39 | layout_models = load_layout_models() 40 | 41 | page_results = extract_tables(images, highres_images, text_lines, det_models, layout_models, rec_models, skip_detection=skip_detection, detect_boxes=detect_cell_boxes) 42 | 43 | out_json = defaultdict(list) 44 | for name, pnum, result in zip(names, pnums, page_results): 45 | for i in range(result.total): 46 | page_cells = result.cells[i] 47 | page_rc = result.rows_cols[i] 48 | img = result.table_imgs[i] 49 | 50 | base_path = os.path.join(out_folder, name) 51 | os.makedirs(base_path, exist_ok=True) 52 | 53 | formatted_result, ext = formatter(format, page_cells) 54 | base_name = f"page{pnum}_table{i}" 55 | with open(os.path.join(base_path, f"{base_name}.{ext}"), "w+", encoding="utf-8") as f: 56 | f.write(formatted_result) 57 | 58 | img.save(os.path.join(base_path, f"{base_name}.png")) 59 | 60 | res = { 61 | "cells": [c.model_dump() for c in page_cells], 62 | "rows": [r.model_dump() for r in page_rc.rows], 63 | "cols": [c.model_dump() for c in page_rc.cols], 64 | "bbox": result.bboxes[i].bbox, 65 | "image_bbox": result.image_bboxes[i].bbox, 66 | "pnum": pnum, 67 | "tnum": i 68 | } 69 | out_json[name].append(res) 70 | 71 | if save_debug_images: 72 | boxes = [l.bbox for l in page_cells] 73 | labels = [l.label for l in page_cells] 74 | bbox_image = draw_bboxes_on_image(boxes, copy.deepcopy(img), labels=labels, label_font_size=20) 75 | bbox_image.save(os.path.join(base_path, f"{base_name}_cells.png")) 76 | 77 | rows = [l.bbox for l in page_rc.rows] 78 | cols = [l.bbox for l in page_rc.cols] 79 | row_labels = [f"Row {l.row_id}" for l in page_rc.rows] 80 | col_labels = [f"Col {l.col_id}" for l in page_rc.cols] 81 | 82 | rc_image = copy.deepcopy(img) 83 | rc_image = draw_bboxes_on_image(rows, rc_image, labels=row_labels, label_font_size=20, color="blue") 84 | rc_image = draw_bboxes_on_image(cols, rc_image, labels=col_labels, label_font_size=20, color="red") 85 | rc_image.save(os.path.join(base_path, f"{base_name}_rc.png")) 86 | 87 | if save_json: 88 | with open(os.path.join(out_folder, "result.json"), "w+", encoding="utf-8") as f: 89 | json.dump(out_json, f, ensure_ascii=False) 90 | 91 | 92 | if __name__ == "__main__": 93 | main() 94 | -------------------------------------------------------------------------------- /pyproject.toml: -------------------------------------------------------------------------------- 1 | [tool.poetry] 2 | name = "tabled-pdf" 3 | version = "0.2.0" 4 | description = "Detect and recognize tables in PDFs and images." 5 | authors = ["Vik Paruchuri "] 6 | readme = "README.md" 7 | license = "GPL-3.0-or-later" 8 | repository = "https://github.com/VikParuchuri/tabled" 9 | keywords = ["table", "table-recognition", "ocr", "pdf"] 10 | packages = [ 11 | {include = "tabled"} 12 | ] 13 | include = [ 14 | "extract.py", 15 | "table_app.py", 16 | "run_table_app.py", 17 | ] 18 | 19 | [tool.poetry.dependencies] 20 | python = "^3.10" 21 | surya-ocr = "~0.8.0" 22 | click = "^8.1.7" 23 | pypdfium2 = "^4.30.0" 24 | pydantic-settings = "^2.5.2" 25 | pydantic = "^2.9.2" 26 | python-dotenv = "^1.0.1" 27 | tabulate = "^0.9.0" 28 | scikit-learn = "^1.5.2" 29 | 30 | [tool.poetry.group.dev.dependencies] 31 | jupyter = "^1.1.1" 32 | datasets = "^3.0.1" 33 | streamlit = "^1.39.0" 34 | rapidfuzz = "^3.10.0" 35 | 36 | [tool.poetry.scripts] 37 | tabled_gui = "run_table_app:run_app" 38 | tabled = "extract:main" 39 | 40 | [build-system] 41 | requires = ["poetry-core"] 42 | build-backend = "poetry.core.masonry.api" 43 | -------------------------------------------------------------------------------- /run_table_app.py: -------------------------------------------------------------------------------- 1 | import subprocess 2 | import os 3 | 4 | 5 | def run_app(): 6 | cur_dir = os.path.dirname(os.path.abspath(__file__)) 7 | ocr_app_path = os.path.join(cur_dir, "table_app.py") 8 | cmd = ["streamlit", "run", ocr_app_path] 9 | subprocess.run(cmd, env={**os.environ, "IN_STREAMLIT": "true"}) 10 | 11 | 12 | if __name__ == "__main__": 13 | run_app() -------------------------------------------------------------------------------- /scripts/verify_benchmark_scores.py: -------------------------------------------------------------------------------- 1 | import json 2 | import argparse 3 | 4 | import click 5 | 6 | 7 | @click.command() 8 | @click.argument("file_path", type=str) 9 | def verify_table_scores(file_path): 10 | with open(file_path, 'r') as file: 11 | data = json.load(file) 12 | 13 | avg = sum([r["score"] for r in data]) / len(data) 14 | if avg < 0.7: 15 | raise ValueError("Average score is below the required threshold of 0.7") 16 | 17 | 18 | if __name__ == "__main__": 19 | verify_table_scores() 20 | -------------------------------------------------------------------------------- /static/images/table_example.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/VikParuchuri/tabled/4a049e83cd14471a59d557adf4329d06cc5b4778/static/images/table_example.png -------------------------------------------------------------------------------- /table_app.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | from tabled.assignment import assign_rows_columns 4 | from tabled.fileinput import load_pdfs_images 5 | from tabled.formats.markdown import markdown_format 6 | from tabled.inference.detection import detect_tables 7 | from tabled.inference.recognition import get_cells, recognize_tables 8 | 9 | os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1" 10 | os.environ["IN_STREAMLIT"] = "true" 11 | import pypdfium2 12 | 13 | import io 14 | import tempfile 15 | from PIL import Image 16 | 17 | import streamlit as st 18 | from tabled.inference.models import load_detection_models, load_recognition_models, load_layout_models 19 | 20 | 21 | @st.cache_resource() 22 | def load_models(): 23 | return load_detection_models(), load_recognition_models(), load_layout_models() 24 | 25 | 26 | def run_table_rec(image, highres_image, text_line, models, skip_detection=False, detect_boxes=False): 27 | if not skip_detection: 28 | table_imgs, table_bboxes, _ = detect_tables([image], [highres_image], models[2]) 29 | else: 30 | table_imgs = [highres_image] 31 | table_bboxes = [[0, 0, highres_image.size[0], highres_image.size[1]]] 32 | 33 | table_text_lines = [text_line] * len(table_imgs) 34 | highres_image_sizes = [highres_image.size] * len(table_imgs) 35 | cells, needs_ocr = get_cells(table_imgs, table_bboxes, highres_image_sizes, table_text_lines, models[0], detect_boxes=detect_boxes) 36 | 37 | table_rec = recognize_tables(table_imgs, cells, needs_ocr, models[1]) 38 | cells = [assign_rows_columns(tr, im_size) for tr, im_size in zip(table_rec, highres_image_sizes)] 39 | 40 | out_data = [] 41 | for idx, (cell, pred, table_img) in enumerate(zip(cells, table_rec, table_imgs)): 42 | md = markdown_format(cell) 43 | out_data.append((md, table_img)) 44 | return out_data 45 | 46 | 47 | def open_pdf(pdf_file): 48 | stream = io.BytesIO(pdf_file.getvalue()) 49 | return pypdfium2.PdfDocument(stream) 50 | 51 | 52 | @st.cache_data() 53 | def get_page_image(pdf_file, page_num, dpi=96): 54 | doc = open_pdf(pdf_file) 55 | renderer = doc.render( 56 | pypdfium2.PdfBitmap.to_pil, 57 | page_indices=[page_num - 1], 58 | scale=dpi / 72, 59 | ) 60 | png = list(renderer)[0] 61 | png_image = png.convert("RGB") 62 | return png_image 63 | 64 | 65 | @st.cache_data() 66 | def page_count(pdf_file): 67 | doc = open_pdf(pdf_file) 68 | return len(doc) 69 | 70 | 71 | st.set_page_config(layout="wide") 72 | 73 | models = load_models() 74 | 75 | 76 | st.markdown(""" 77 | # Tabled Demo 78 | 79 | This app will let you try tabled, a table detection and recognition model. It will detect and recognize the tables. 80 | 81 | Find the project [here](https://github.com/VikParuchuri/tabled). 82 | """) 83 | 84 | in_file = st.sidebar.file_uploader("PDF file or image:", type=["pdf", "png", "jpg", "jpeg", "gif", "webp"]) 85 | skip_detection = st.sidebar.checkbox("Skip table detection", help="Use this if tables are already cropped (the whole PDF page or image is a table)", value=False) 86 | detect_boxes = st.sidebar.checkbox("Detect cell boxes", help="Detect table cell boxes vs extract from PDF. Will also run OCR.", value=False) 87 | 88 | if in_file is None: 89 | st.stop() 90 | 91 | filetype = in_file.type 92 | col = st.columns(1)[0] 93 | container = col.container() 94 | 95 | if "pdf" in filetype: 96 | page_count = page_count(in_file) 97 | page_number = st.sidebar.number_input(f"Page number out of {page_count}:", min_value=1, value=1, 98 | max_value=page_count) 99 | 100 | pil_image = get_page_image(in_file, page_number, 96) 101 | else: 102 | pil_image = Image.open(in_file).convert("RGB") 103 | pil_image_highres = pil_image 104 | page_number = 1 105 | 106 | with col: 107 | st.image(pil_image, caption="PDF file (preview)", use_container_width=True) 108 | 109 | run_marker = st.sidebar.button("Run Tabled") 110 | 111 | if not run_marker: 112 | st.stop() 113 | 114 | # Run Tabled 115 | file_ext = in_file.name.rsplit(".")[-1] 116 | with tempfile.NamedTemporaryFile(suffix=file_ext) as temp_input: 117 | temp_input.write(in_file.getvalue()) 118 | temp_input.seek(0) 119 | filename = temp_input.name 120 | images, highres_images, names, text_lines = load_pdfs_images(filename, max_pages=1, start_page=page_number - 1) 121 | out_data = run_table_rec(images[0], highres_images[0], text_lines[0], models, skip_detection=skip_detection, detect_boxes=detect_boxes) 122 | 123 | for idx, (md, table_img) in enumerate(out_data): 124 | container.markdown(f"## Table {idx}") 125 | container.image(table_img, caption=f"Table {idx}", use_container_width=True) 126 | container.markdown(md) 127 | container.code(md) 128 | container.divider() 129 | 130 | -------------------------------------------------------------------------------- /tabled/assignment.py: -------------------------------------------------------------------------------- 1 | from typing import List 2 | 3 | import numpy as np 4 | from surya.schema import TableResult, Bbox 5 | 6 | from tabled.heuristics import heuristic_layout 7 | from tabled.schema import SpanTableCell 8 | 9 | 10 | def is_rotated(rows, cols): 11 | # Determine if the table is rotated by looking at row and column width / height ratios 12 | # Rows should have a >1 ratio, cols <1 13 | widths = sum([r.width for r in rows]) 14 | heights = sum([c.height for c in rows]) + 1 15 | r_ratio = widths / heights 16 | 17 | widths = sum([c.width for c in cols]) 18 | heights = sum([r.height for r in cols]) + 1 19 | c_ratio = widths / heights 20 | 21 | return r_ratio * 2 < c_ratio 22 | 23 | 24 | def overlapper_idxs(rows, field, thresh=.3): 25 | overlapper_rows = set() 26 | for row in rows: 27 | row_id = getattr(row, field) 28 | if row_id in overlapper_rows: 29 | continue 30 | 31 | for row2 in rows: 32 | row2_id = getattr(row2, field) 33 | if row2_id == row_id or row2_id in overlapper_rows: 34 | continue 35 | 36 | if row.intersection_pct(row2) > thresh: 37 | i_bigger = row.area > row2.area 38 | overlapper_rows.add(row_id if i_bigger else row2_id) 39 | return overlapper_rows 40 | 41 | 42 | def initial_assignment(detection_result: TableResult, thresh=.5) -> List[SpanTableCell]: 43 | overlapper_rows = overlapper_idxs(detection_result.rows, field="row_id") 44 | overlapper_cols = overlapper_idxs(detection_result.cols, field="col_id") 45 | 46 | cells = [] 47 | for cell in detection_result.cells: 48 | max_intersection = 0 49 | row_pred = None 50 | for row in detection_result.rows: 51 | if row.row_id in overlapper_rows: 52 | continue 53 | 54 | intersection_pct = Bbox(bbox=cell.bbox).intersection_pct(row) 55 | if intersection_pct > max_intersection and intersection_pct > thresh: 56 | max_intersection = intersection_pct 57 | row_pred = row.row_id 58 | 59 | max_intersection = 0 60 | col_pred = None 61 | for col in detection_result.cols: 62 | if col.col_id in overlapper_cols: 63 | continue 64 | 65 | intersection_pct = Bbox(bbox=cell.bbox).intersection_pct(col) 66 | if intersection_pct > max_intersection and intersection_pct > thresh: 67 | max_intersection = intersection_pct 68 | col_pred = col.col_id 69 | 70 | cells.append( 71 | SpanTableCell( 72 | bbox=cell.bbox, 73 | text=cell.text, 74 | row_ids=[row_pred], 75 | col_ids=[col_pred] 76 | ) 77 | ) 78 | return cells 79 | 80 | 81 | def assign_overlappers(cells: List[SpanTableCell], detection_result: TableResult, thresh=.5): 82 | overlapper_rows = overlapper_idxs(detection_result.rows, field="row_id") 83 | overlapper_cols = overlapper_idxs(detection_result.cols, field="col_id") 84 | 85 | for cell in cells: 86 | max_intersection = 0 87 | row_pred = None 88 | for row in detection_result.rows: 89 | if row.row_id not in overlapper_rows: 90 | continue 91 | 92 | intersection_pct = Bbox(bbox=cell.bbox).intersection_pct(row) 93 | if intersection_pct > max_intersection and intersection_pct > thresh: 94 | max_intersection = intersection_pct 95 | row_pred = row.row_id 96 | 97 | max_intersection = 0 98 | col_pred = None 99 | for col in detection_result.cols: 100 | if col.col_id not in overlapper_cols: 101 | continue 102 | 103 | intersection_pct = Bbox(bbox=cell.bbox).intersection_pct(col) 104 | if intersection_pct > max_intersection and intersection_pct > thresh: 105 | max_intersection = intersection_pct 106 | col_pred = col.col_id 107 | 108 | if cell.row_ids[0] is None: 109 | cell.row_ids = [row_pred] 110 | if cell.col_ids[0] is None: 111 | cell.col_ids = [col_pred] 112 | 113 | 114 | def assign_unassigned(table_cells: list, detection_result: TableResult): 115 | rotated = is_rotated(detection_result.rows, detection_result.cols) 116 | for cell in table_cells: 117 | if cell.row_ids[0] is None: 118 | closest_row = None 119 | min_dist = None 120 | for row in detection_result.rows: 121 | if rotated: 122 | dist = cell.center_x_distance(row) 123 | else: 124 | dist = cell.center_y_distance(row) 125 | 126 | if min_dist is None or dist < min_dist: 127 | closest_row = row.row_id 128 | min_dist = dist 129 | cell.row_ids = [closest_row] 130 | 131 | if cell.col_ids[0] is None: 132 | closest_col = None 133 | min_dist = None 134 | for col in detection_result.cols: 135 | if rotated: 136 | dist = cell.center_y_distance(col) 137 | else: 138 | dist = cell.center_x_distance(col) 139 | 140 | if min_dist is None or dist < min_dist: 141 | closest_col = col.col_id 142 | min_dist = dist 143 | 144 | cell.col_ids = [closest_col] 145 | 146 | 147 | def handle_rowcol_spans(table_cells: list, detection_result: TableResult, thresh=.25): 148 | rotated = is_rotated(detection_result.rows, detection_result.cols) 149 | for cell in table_cells: 150 | for c in detection_result.cols: 151 | col_intersect_pct = cell.intersection_y_pct(c) if rotated else cell.intersection_x_pct(c) 152 | other_cell_exists = len([tc for tc in table_cells if tc.col_ids[0] == c.col_id and tc.row_ids[0] == cell.row_ids[0]]) > 0 153 | if col_intersect_pct > thresh and not other_cell_exists: 154 | cell.col_ids.append(c.col_id) 155 | else: 156 | break 157 | # Assign to first column header appears in 158 | cell.col_ids = sorted(cell.col_ids) 159 | 160 | for cell in table_cells: 161 | for r in detection_result.rows: 162 | row_intersect_pct = cell.intersection_x_pct(r) if rotated else cell.intersection_y_pct(r) 163 | other_cell_exists = len([tc for tc in table_cells if tc.row_ids[0] == r.row_id and tc.col_ids[0] == cell.col_ids[0]]) > 0 164 | if row_intersect_pct > thresh and not other_cell_exists: 165 | cell.row_ids.append(r.row_id) 166 | else: 167 | break 168 | 169 | 170 | def merge_multiline_rows(detection_result: TableResult, table_cells: List[SpanTableCell]): 171 | def find_row_gap(r1, r2): 172 | return min([abs(r1.bbox[1] - r2.bbox[3]), abs(r2.bbox[1] - r1.bbox[3])]) 173 | 174 | all_cols = set([tc.col_ids[0] for tc in table_cells]) 175 | if len(all_cols) == 0: 176 | return 177 | 178 | merged_pairs = [] 179 | row_gaps = [ 180 | find_row_gap(r, r2) 181 | for r, r2 in zip(detection_result.rows, detection_result.rows[1:]) 182 | ] 183 | if len(row_gaps) == 0: 184 | return 185 | 186 | gap_thresh = np.median(row_gaps) 187 | 188 | for idx, row in enumerate(detection_result.rows[1:]): 189 | prev_row = detection_result.rows[idx - 1] 190 | gap = find_row_gap(prev_row, row) 191 | 192 | # Ensure the gap between r2 and r1 is small 193 | if gap > gap_thresh: 194 | continue 195 | 196 | r1_cells = [tc for tc in table_cells if tc.row_ids[0] == prev_row.row_id] 197 | r2_cells = [tc for tc in table_cells if tc.row_ids[0] == row.row_id] 198 | r1_cols = set([tc.col_ids[0] for tc in r1_cells]) 199 | r2_cols = set([tc.col_ids[0] for tc in r2_cells]) 200 | 201 | # Ensure all columns in r2 are in r1 202 | if len(r2_cols - r1_cols) > 0: 203 | continue 204 | 205 | # Ensure r2 has mostly blank cells 206 | if len(r2_cols) / len(all_cols) > .5: 207 | continue 208 | 209 | merged_pairs.append((idx - 1, idx)) 210 | 211 | to_remove = set() 212 | for r1_idx, r2_idx in merged_pairs: 213 | detection_result.rows[r1_idx].bbox = [ 214 | min(detection_result.rows[r1_idx].bbox[0], detection_result.rows[r2_idx].bbox[0]), 215 | min(detection_result.rows[r1_idx].bbox[1], detection_result.rows[r2_idx].bbox[1]), 216 | max(detection_result.rows[r1_idx].bbox[2], detection_result.rows[r2_idx].bbox[2]), 217 | max(detection_result.rows[r1_idx].bbox[3], detection_result.rows[r2_idx].bbox[3]) 218 | ] 219 | to_remove.add(r2_idx) 220 | 221 | new_rows = [] 222 | row_counter = 0 223 | for idx, row in enumerate(detection_result.rows): 224 | if idx not in to_remove: 225 | row.row_id = row_counter 226 | new_rows.append(row) 227 | row_counter += 1 228 | detection_result.rows = new_rows 229 | 230 | 231 | def assign_rows_columns(detection_result: TableResult, image_size: list, heuristic_thresh=.6) -> List[SpanTableCell]: 232 | table_cells = initial_assignment(detection_result) 233 | merge_multiline_rows(detection_result, table_cells) 234 | table_cells = initial_assignment(detection_result) 235 | assign_overlappers(table_cells, detection_result) 236 | total_unassigned = len([tc for tc in table_cells if tc.row_ids[0] is None or tc.col_ids[0] is None]) 237 | unassigned_frac = total_unassigned / max(len(table_cells), 1) 238 | 239 | if unassigned_frac > heuristic_thresh: 240 | table_cells = heuristic_layout(table_cells, image_size) 241 | return table_cells 242 | 243 | assign_unassigned(table_cells, detection_result) 244 | handle_rowcol_spans(table_cells, detection_result) 245 | return table_cells 246 | -------------------------------------------------------------------------------- /tabled/extract.py: -------------------------------------------------------------------------------- 1 | from typing import List 2 | 3 | from surya.schema import Bbox 4 | 5 | from tabled.assignment import assign_rows_columns 6 | from tabled.inference.detection import detect_tables 7 | from tabled.inference.recognition import get_cells, recognize_tables 8 | from tabled.schema import ExtractPageResult 9 | 10 | 11 | def extract_tables( 12 | images, 13 | highres_images, 14 | text_lines, 15 | det_models, 16 | layout_models, 17 | rec_models, 18 | skip_detection=False, 19 | detect_boxes=False 20 | ) -> List[ExtractPageResult]: 21 | if not skip_detection: 22 | table_imgs, table_bboxes, table_counts = detect_tables(images, highres_images, layout_models) 23 | else: 24 | table_imgs = highres_images 25 | table_bboxes = [[0, 0, img.size[0], img.size[1]] for img in highres_images] 26 | table_counts = [1] * len(highres_images) 27 | 28 | table_text_lines = [] 29 | highres_image_sizes = [] 30 | for i, tc in enumerate(table_counts): 31 | table_text_lines.extend([text_lines[i]] * tc) 32 | highres_image_sizes.extend([highres_images[i].size] * tc) 33 | 34 | cells, needs_ocr = get_cells(table_imgs, table_bboxes, highres_image_sizes, table_text_lines, det_models, detect_boxes=detect_boxes) 35 | 36 | table_rec = recognize_tables(table_imgs, cells, needs_ocr, rec_models) 37 | cells = [assign_rows_columns(tr, im_size) for tr, im_size in zip(table_rec, highres_image_sizes)] 38 | 39 | results = [] 40 | counter = 0 41 | for count in table_counts: 42 | page_start = counter 43 | page_end = counter + count 44 | results.append(ExtractPageResult( 45 | table_imgs=table_imgs[page_start:page_end], 46 | cells=cells[page_start:page_end], 47 | rows_cols=table_rec[page_start:page_end], 48 | bboxes=[Bbox(bbox=b) for b in table_bboxes[page_start:page_end]], 49 | image_bboxes=[Bbox(bbox=[0, 0, size[0], size[1]]) for size in highres_image_sizes[page_start:page_end]] 50 | )) 51 | counter += count 52 | 53 | assert len(results) == len(images) 54 | return results 55 | -------------------------------------------------------------------------------- /tabled/fileinput.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | from surya.input.load import load_from_folder, load_from_file 4 | from surya.settings import settings as surya_settings 5 | 6 | 7 | def load_pdfs_images(input_path, max_pages=None, start_page=None): 8 | if os.path.isdir(input_path): 9 | images, _, _ = load_from_folder(input_path, max_pages, start_page=start_page) 10 | highres_images, names, text_lines = load_from_folder(input_path, max_pages, dpi=surya_settings.IMAGE_DPI_HIGHRES, 11 | load_text_lines=True, start_page=start_page) 12 | else: 13 | images, _, _ = load_from_file(input_path, max_pages, start_page=start_page) 14 | highres_images, names, text_lines = load_from_file(input_path, max_pages, dpi=surya_settings.IMAGE_DPI_HIGHRES, 15 | load_text_lines=True, start_page=start_page) 16 | 17 | return images, highres_images, names, text_lines -------------------------------------------------------------------------------- /tabled/formats/__init__.py: -------------------------------------------------------------------------------- 1 | from tabled.formats.csv import csv_format 2 | from tabled.formats.html import html_format 3 | from tabled.formats.markdown import markdown_format 4 | 5 | 6 | def formatter(format, page_cells): 7 | if format == "csv": 8 | return csv_format(page_cells), "csv" 9 | elif format == "markdown": 10 | return markdown_format(page_cells), "md" 11 | elif format == "html": 12 | return html_format(page_cells), "html" 13 | else: 14 | raise ValueError(f"Invalid format: {format}") -------------------------------------------------------------------------------- /tabled/formats/common.py: -------------------------------------------------------------------------------- 1 | import re 2 | from typing import List 3 | 4 | from tabled.schema import SpanTableCell 5 | 6 | 7 | def sort_within_cell(cells, tolerance=5): 8 | vertical_groups = {} 9 | for i, cell in enumerate(cells): 10 | group_key = round((cell.bbox[1] + cell.bbox[3]) / 2 / tolerance) 11 | if group_key not in vertical_groups: 12 | vertical_groups[group_key] = [] 13 | vertical_groups[group_key].append((i, cell.bbox[0])) 14 | 15 | # Sort each group horizontally and flatten the groups into a single list 16 | sorted_cell_idxs = [] 17 | for _, group in sorted(vertical_groups.items()): 18 | sorted_group = sorted(group, key=lambda x: x[1]) 19 | sorted_cell_idxs.extend([idx for idx, _ in sorted_group]) 20 | 21 | cell_order = [sorted_cell_idxs.index(i) for i in range(len(sorted_cell_idxs))] 22 | return cell_order 23 | 24 | 25 | def sort_cells(cells: List[SpanTableCell]): 26 | cell_order = sort_within_cell(cells) 27 | for i, cell in enumerate(cells): 28 | cell.order = cell_order[i] 29 | cells.sort(key=lambda x: (x.row_ids[0], x.col_ids[0], x.order)) 30 | return cells 31 | 32 | 33 | def replace_dots(text): 34 | dot_pattern = re.compile(r'(\s*\.\s*){4,}') 35 | dot_multiline_pattern = re.compile(r'.*(\s*\.\s*){4,}.*', re.DOTALL) 36 | 37 | if dot_multiline_pattern.match(text): 38 | text = dot_pattern.sub(' ', text) 39 | return text 40 | 41 | 42 | def replace_newlines(text): 43 | # Replace all newlines 44 | newline_pattern = re.compile(r'[\r\n]+') 45 | return newline_pattern.sub(' ', text).strip() 46 | -------------------------------------------------------------------------------- /tabled/formats/csv.py: -------------------------------------------------------------------------------- 1 | from typing import List 2 | 3 | from tabulate import tabulate 4 | 5 | from tabled.formats.common import sort_cells, replace_dots, replace_newlines 6 | from tabled.schema import SpanTableCell 7 | import csv 8 | import io 9 | 10 | 11 | def replace_all(text): 12 | return replace_newlines(replace_dots(text)) 13 | 14 | 15 | def csv_format(cells: List[SpanTableCell]): 16 | cells = sort_cells(cells) 17 | unique_rows = set([cell.row_ids[0] for cell in cells]) 18 | unique_cols = set([cell.col_ids[0] for cell in cells]) 19 | buff = io.StringIO() 20 | writer = csv.writer(buff) 21 | for row in unique_rows: 22 | text_row = [] 23 | for col in unique_cols: 24 | cell = " ".join([cell.text for cell in cells if cell.row_ids[0] == row and cell.col_ids[0] == col]) 25 | cell = replace_all(cell) 26 | text_row.append(cell) 27 | writer.writerow(text_row) 28 | 29 | csv_str = buff.getvalue() 30 | return csv_str 31 | -------------------------------------------------------------------------------- /tabled/formats/html.py: -------------------------------------------------------------------------------- 1 | from typing import List 2 | 3 | from tabulate import tabulate 4 | 5 | from tabled.formats.common import sort_cells, replace_dots, replace_newlines 6 | from tabled.schema import SpanTableCell 7 | 8 | 9 | def replace_all(text): 10 | return replace_newlines(replace_dots(text)) 11 | 12 | 13 | def html_format(cells: List[SpanTableCell]): 14 | md_rows = [] 15 | cells = sort_cells(cells) 16 | unique_rows = set([cell.row_ids[0] for cell in cells]) 17 | unique_cols = set([cell.col_ids[0] for cell in cells]) 18 | for row in unique_rows: 19 | md_row = [] 20 | for col in unique_cols: 21 | cell = " ".join([cell.text for cell in cells if cell.row_ids[0] == row and cell.col_ids[0] == col]) 22 | cell = replace_all(cell) 23 | md_row.append(cell) 24 | md_rows.append(md_row) 25 | 26 | headers = "firstrow" if len(cells) > 1 else "" 27 | md = tabulate(md_rows, headers=headers, tablefmt="html", disable_numparse=True) 28 | return md 29 | -------------------------------------------------------------------------------- /tabled/formats/markdown.py: -------------------------------------------------------------------------------- 1 | from typing import List 2 | 3 | from tabulate import tabulate 4 | 5 | from tabled.formats.common import sort_cells, replace_dots, replace_newlines 6 | from tabled.schema import SpanTableCell 7 | 8 | 9 | def replace_special_chars(text): 10 | return text.replace("|", "\\|").replace("-", "\\-") 11 | 12 | 13 | def replace_all(text): 14 | return replace_special_chars(replace_newlines(replace_dots(text))) 15 | 16 | 17 | def markdown_format(cells: List[SpanTableCell]): 18 | md_rows = [] 19 | cells = sort_cells(cells) 20 | unique_rows = set([cell.row_ids[0] for cell in cells]) 21 | unique_cols = set([cell.col_ids[0] for cell in cells]) 22 | for row in unique_rows: 23 | md_row = [] 24 | for col in unique_cols: 25 | cell = " ".join([cell.text for cell in cells if cell.row_ids[0] == row and cell.col_ids[0] == col]) 26 | cell = replace_all(cell) 27 | md_row.append(cell) 28 | md_rows.append(md_row) 29 | 30 | md = tabulate(md_rows, headers="firstrow", tablefmt="github", disable_numparse=True) 31 | return md 32 | -------------------------------------------------------------------------------- /tabled/heuristics/__init__.py: -------------------------------------------------------------------------------- 1 | from typing import List 2 | 3 | from tabled.heuristics.cells import assign_cells_to_columns 4 | from tabled.schema import SpanTableCell 5 | 6 | 7 | def heuristic_layout(table_cells: List[SpanTableCell], page_size, row_tol=.01) -> List[SpanTableCell]: 8 | table_rows = [] 9 | table_row = [] 10 | y_top = None 11 | y_bottom = None 12 | for cell in table_cells: 13 | normed_y_start = cell.bbox[1] / page_size[1] 14 | normed_y_end = cell.bbox[3] / page_size[1] 15 | 16 | if y_top is None: 17 | y_top = normed_y_start 18 | if y_bottom is None: 19 | y_bottom = normed_y_end 20 | 21 | y_dist = min(abs(normed_y_start - y_bottom), abs(normed_y_end - y_bottom)) 22 | if y_dist < row_tol: 23 | table_row.append(cell) 24 | else: 25 | # New row 26 | if len(table_row) > 0: 27 | table_rows.append(table_row) 28 | table_row = [cell] 29 | y_top = normed_y_start 30 | y_bottom = normed_y_end 31 | if len(table_row) > 0: 32 | table_rows.append(table_row) 33 | 34 | return assign_cells_to_columns(table_rows, page_size) -------------------------------------------------------------------------------- /tabled/heuristics/cells.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from sklearn.cluster import DBSCAN 3 | 4 | 5 | def cluster_coords(coords, row_count): 6 | if len(coords) == 0: 7 | return [] 8 | coords = np.array(sorted(set(coords))).reshape(-1, 1) 9 | 10 | clustering = DBSCAN(eps=.01, min_samples=max(2, row_count // 4)).fit(coords) 11 | clusters = clustering.labels_ 12 | 13 | separators = [] 14 | for label in set(clusters): 15 | clustered_points = coords[clusters == label] 16 | separators.append(np.mean(clustered_points)) 17 | 18 | separators = sorted(separators) 19 | return separators 20 | 21 | 22 | def find_column_separators(rows, page_size, round_factor=.002, min_count=1): 23 | left_edges = [] 24 | right_edges = [] 25 | centers = [] 26 | 27 | boxes = [c.bbox for r in rows for c in r] 28 | 29 | for cell in boxes: 30 | ncell = [cell[0] / page_size[0], cell[1] / page_size[1], cell[2] / page_size[0], cell[3] / page_size[1]] 31 | left_edges.append(ncell[0] / round_factor * round_factor) 32 | right_edges.append(ncell[2] / round_factor * round_factor) 33 | centers.append((ncell[0] + ncell[2]) / 2 * round_factor / round_factor) 34 | 35 | left_edges = [l for l in left_edges if left_edges.count(l) > min_count] 36 | right_edges = [r for r in right_edges if right_edges.count(r) > min_count] 37 | centers = [c for c in centers if centers.count(c) > min_count] 38 | 39 | sorted_left = cluster_coords(left_edges, len(rows)) 40 | sorted_right = cluster_coords(right_edges, len(rows)) 41 | sorted_center = cluster_coords(centers, len(rows)) 42 | 43 | # Find list with minimum length 44 | separators = max([sorted_left, sorted_right, sorted_center], key=len) 45 | separators.append(1) 46 | separators.insert(0, 0) 47 | return separators 48 | 49 | 50 | def assign_cells_to_columns(rows, page_size, round_factor=.002, tolerance=.01): 51 | separators = find_column_separators(rows, page_size, round_factor=round_factor) 52 | additional_column_index = 0 53 | row_dicts = [] 54 | 55 | for row in rows: 56 | new_row = {} 57 | last_col_index = -1 58 | for cell in row: 59 | left_edge = cell.bbox[0] / page_size[0] 60 | column_index = -1 61 | for i, separator in enumerate(separators): 62 | if left_edge - tolerance < separator and last_col_index < i: 63 | column_index = i 64 | break 65 | if column_index == -1: 66 | column_index = len(separators) + additional_column_index 67 | additional_column_index += 1 68 | new_row[column_index] = cell 69 | last_col_index = column_index 70 | additional_column_index = 0 71 | row_dicts.append(new_row) 72 | 73 | cells = [] 74 | for row_idx, row in enumerate(row_dicts): 75 | column = 0 76 | for col_idx in sorted(row.keys()): 77 | cell = row[col_idx] 78 | cell.row_ids = [row_idx] 79 | cell.col_ids = [column] 80 | cells.append(cell) 81 | column += 1 82 | 83 | return cells -------------------------------------------------------------------------------- /tabled/inference/detection.py: -------------------------------------------------------------------------------- 1 | from surya.layout import batch_layout_detection 2 | from surya.postprocessing.util import rescale_bbox 3 | from surya.schema import Bbox 4 | 5 | from tabled.settings import settings 6 | 7 | 8 | def merge_boxes(box1, box2): 9 | return [min(box1[0], box2[0]), min(box1[1], box2[1]), max(box1[2], box2[2]), max(box1[3], box2[3])] 10 | 11 | 12 | def merge_tables(page_table_boxes): 13 | # Merge tables that are next to each other 14 | expansion_factor = 1.02 15 | shrink_factor = .98 16 | ignore_boxes = set() 17 | for i in range(len(page_table_boxes)): 18 | if i in ignore_boxes: 19 | continue 20 | for j in range(i + 1, len(page_table_boxes)): 21 | if j in ignore_boxes: 22 | continue 23 | expanded_box1 = [page_table_boxes[i][0] * shrink_factor, page_table_boxes[i][1], 24 | page_table_boxes[i][2] * expansion_factor, page_table_boxes[i][3]] 25 | expanded_box2 = [page_table_boxes[j][0] * shrink_factor, page_table_boxes[j][1], 26 | page_table_boxes[j][2] * expansion_factor, page_table_boxes[j][3]] 27 | if Bbox(bbox=expanded_box1).intersection_pct(Bbox(bbox=expanded_box2)) > 0: 28 | page_table_boxes[i] = merge_boxes(page_table_boxes[i], page_table_boxes[j]) 29 | ignore_boxes.add(j) 30 | 31 | return [b for i, b in enumerate(page_table_boxes) if i not in ignore_boxes] 32 | 33 | 34 | def detect_tables(images, highres_images, models, layout_batch_size=settings.LAYOUT_BATCH_SIZE): 35 | layout_model, layout_processor = models 36 | layout_predictions = batch_layout_detection(images, layout_model, layout_processor, batch_size=layout_batch_size) 37 | 38 | table_imgs = [] 39 | table_counts = [] 40 | table_bboxes = [] 41 | 42 | for layout_pred, img, highres_img in zip(layout_predictions, images, highres_images): 43 | # The bbox for the entire table 44 | bbox = [l.bbox for l in layout_pred.bboxes if l.label == "Table"] 45 | 46 | if len(bbox) == 0: 47 | table_counts.append(0) 48 | continue 49 | 50 | page_table_imgs = [] 51 | highres_bbox = [] 52 | 53 | # Merge tables that are next to each other 54 | bbox = merge_tables(bbox) 55 | 56 | # Number of tables per page 57 | table_counts.append(len(bbox)) 58 | 59 | for bb in bbox: 60 | highres_bb = rescale_bbox(bb, img.size, highres_img.size) 61 | page_table_imgs.append(highres_img.crop(highres_bb)) 62 | highres_bbox.append(highres_bb) 63 | 64 | table_imgs.extend(page_table_imgs) 65 | table_bboxes.extend(highres_bbox) 66 | 67 | return table_imgs, table_bboxes, table_counts -------------------------------------------------------------------------------- /tabled/inference/models.py: -------------------------------------------------------------------------------- 1 | from surya.model.detection.model import load_model as load_det_model, load_processor as load_det_processor 2 | from surya.model.layout.model import load_model as load_layout_model 3 | from surya.model.layout.processor import load_processor as load_layout_processor 4 | from surya.model.recognition.model import load_model as load_rec_model 5 | from surya.model.recognition.processor import load_processor as load_rec_processor 6 | from surya.model.table_rec.model import load_model as load_table_rec_model 7 | from surya.model.table_rec.processor import load_processor as load_table_rec_processor 8 | 9 | 10 | def load_detection_models(): 11 | detection_model = load_det_model() 12 | detection_processor = load_det_processor() 13 | return detection_model, detection_processor 14 | 15 | 16 | def load_recognition_models(): 17 | table_rec_model = load_table_rec_model() 18 | table_rec_processor = load_table_rec_processor() 19 | rec_model = load_rec_model() 20 | rec_processor = load_rec_processor() 21 | return table_rec_model, table_rec_processor, rec_model, rec_processor 22 | 23 | def load_layout_models(): 24 | layout_model = load_layout_model() 25 | layout_processor = load_layout_processor() 26 | return layout_model, layout_processor -------------------------------------------------------------------------------- /tabled/inference/recognition.py: -------------------------------------------------------------------------------- 1 | from typing import List 2 | 3 | from surya.detection import batch_text_detection 4 | from surya.input.pdflines import get_table_blocks 5 | from surya.ocr import run_recognition 6 | from surya.schema import TableResult 7 | from surya.tables import batch_table_recognition 8 | 9 | from tabled.settings import settings 10 | 11 | 12 | def get_cells(table_imgs, table_bboxes, image_sizes, text_lines, models, detect_boxes=False, detector_batch_size=settings.DETECTOR_BATCH_SIZE): 13 | det_model, det_processor = models 14 | table_cells = [] 15 | needs_ocr = [] 16 | 17 | to_inference_idxs = [] 18 | for idx, (highres_bbox, text_line, image_size) in enumerate(zip(table_bboxes, text_lines, image_sizes)): 19 | # The text cells inside each table 20 | table_blocks = get_table_blocks([highres_bbox], text_line, image_size)[0] if text_line is not None else None 21 | 22 | if text_line is None or detect_boxes or len(table_blocks) == 0: 23 | to_inference_idxs.append(idx) 24 | table_cells.append(None) 25 | needs_ocr.append(True) 26 | else: 27 | table_cells.append(table_blocks) 28 | needs_ocr.append(False) 29 | 30 | # Inference tables that need it 31 | if len(to_inference_idxs) > 0: 32 | det_results = batch_text_detection([table_imgs[i] for i in to_inference_idxs], det_model, det_processor, batch_size=detector_batch_size) 33 | for idx, det_result in zip(to_inference_idxs, det_results): 34 | cell_bboxes = [{"bbox": tb.bbox, "text": None} for tb in det_result.bboxes if tb.area > 0] 35 | table_cells[idx] = cell_bboxes 36 | 37 | return table_cells, needs_ocr 38 | 39 | 40 | def recognize_tables(table_imgs, table_cells, needs_ocr: List[bool], models, table_rec_batch_size=settings.TABLE_REC_BATCH_SIZE, ocr_batch_size=settings.RECOGNITION_BATCH_SIZE) -> List[TableResult]: 41 | table_rec_model, table_rec_processor, ocr_model, ocr_processor = models 42 | 43 | if sum(needs_ocr) > 0: 44 | needs_ocr_idx = [idx for idx, needs in enumerate(needs_ocr) if needs] 45 | ocr_images = [img for img, needs in zip(table_imgs, needs_ocr) if needs] 46 | ocr_cells = [[c["bbox"] for c in cells] for cells, needs in zip(table_cells, needs_ocr) if needs] 47 | ocr_langs = [None] * len(ocr_images) 48 | 49 | ocr_predictions = run_recognition(ocr_images, ocr_langs, ocr_model, ocr_processor, bboxes=ocr_cells, batch_size=ocr_batch_size) 50 | 51 | # Assign text to correct spot 52 | for orig_idx, ocr_pred in zip(needs_ocr_idx, ocr_predictions): 53 | for ocr_line, cell in zip(ocr_pred.text_lines, table_cells[orig_idx]): 54 | cell["text"] = ocr_line.text 55 | 56 | table_preds = batch_table_recognition(table_imgs, table_cells, table_rec_model, table_rec_processor, batch_size=table_rec_batch_size) 57 | return table_preds 58 | 59 | 60 | -------------------------------------------------------------------------------- /tabled/schema.py: -------------------------------------------------------------------------------- 1 | from typing import List, Optional, Any 2 | 3 | from pydantic import BaseModel, model_validator 4 | from surya.schema import Bbox, TableResult 5 | 6 | 7 | def str_join(list, char=","): 8 | return char.join([str(x) for x in list]) 9 | 10 | 11 | class SpanTableCell(Bbox): 12 | text: str 13 | row_ids: List[Optional[int]] 14 | col_ids: List[Optional[int]] 15 | order: Optional[int] = None 16 | 17 | def intersection_x_pct(self, other): 18 | if self.width == 0: 19 | return 0 20 | 21 | x_overlap = max(0, min(self.bbox[2], other.bbox[2]) - max(self.bbox[0], other.bbox[0])) 22 | return x_overlap / self.width 23 | 24 | 25 | def intersection_y_pct(self, other): 26 | if self.height == 0: 27 | return 0 28 | 29 | y_overlap = max(0, min(self.bbox[3], other.bbox[3]) - max(self.bbox[1], other.bbox[1])) 30 | return y_overlap / self.height 31 | 32 | @property 33 | def label(self): 34 | return f"{str_join(self.row_ids)}-{str_join(self.col_ids)}" 35 | 36 | def center_x_distance(self, other): 37 | return abs(self.center[0] - other.center[0]) 38 | 39 | def center_y_distance(self, other): 40 | return abs(self.center[1] - other.center[1]) 41 | 42 | 43 | class ExtractPageResult(BaseModel): 44 | cells: List[List[SpanTableCell]] 45 | rows_cols: List[TableResult] 46 | table_imgs: List[Any] 47 | bboxes: List[Bbox] # Bbox of the table 48 | image_bboxes: List[Bbox] # Bbox of the image/page table is inside 49 | 50 | @model_validator(mode="after") 51 | def check_cells(self): 52 | assert len(self.cells) == len(self.table_imgs), "Cells and table images must be the same length" 53 | assert len(self.cells) == len(self.rows_cols), "Cells and rows/cols must be the same length" 54 | return self 55 | 56 | @property 57 | def total(self): 58 | return len(self.cells) 59 | -------------------------------------------------------------------------------- /tabled/settings.py: -------------------------------------------------------------------------------- 1 | from typing import Optional 2 | 3 | from dotenv import find_dotenv 4 | from pydantic_settings import BaseSettings 5 | 6 | 7 | class Settings(BaseSettings): 8 | # General 9 | IN_STREAMLIT: bool = False 10 | TORCH_DEVICE: Optional[str] = None 11 | 12 | # Batch sizes 13 | # See https://github.com/VikParuchuri/surya for default values 14 | ## Table recognition 15 | TABLE_REC_BATCH_SIZE: Optional[int] = None 16 | ## OCR 17 | RECOGNITION_BATCH_SIZE: Optional[int] = None 18 | ## Text detector 19 | DETECTOR_BATCH_SIZE: Optional[int] = None 20 | ## Layout 21 | LAYOUT_BATCH_SIZE: Optional[int] = None 22 | 23 | class Config: 24 | env_file = find_dotenv("local.env") 25 | extra = "ignore" 26 | 27 | 28 | settings = Settings() --------------------------------------------------------------------------------